A new multivariate test formulation: theory, implementation, and applications to genome-scale sequencing and expression

Lei XuApplied Informatics 2016 · DOI

Abstract

Traditional multivariate statistical tests often struggle with the "curse of dimensionality" and a lack of interpretability when applied to complex biological data. This paper proposes a novel formulation for multivariate testing based on a hierarchical lattice taxonomy of properties. By decomposing complex dependencies into independent intrinsic factors and utilizing a "Best First Path" (BFP) search algorithm, this framework identifies specific variable combinations that drive significant results. This approach offers enhanced sensitivity and better interpretability for applications in genome-scale sequencing and expression profile analysis.

The Limitations of Traditional Multivariate Tests

In genomic research, case-control studies are fundamental for identifying differences between healthy and diseased groups. However, standard multivariate tests, such as Hotelling’s $T^2$, face three primary challenges:

  1. Lumping: They treat all variables as a single unit, making it difficult to pinpoint which specific factors contribute to a significant result.
  2. The Curse of Dimensionality: In scenarios with many variables but few samples, estimating covariance matrices becomes mathematically unreliable.
  3. Ignoring Direction: Most tests measure the magnitude of difference but ignore the direction (e.g., collective up-regulation vs. down-regulation), losing valuable biological context.

The Lattice Taxonomy Framework

To address these issues, the proposed method organizes numerous tests into a Lattice Taxonomy of Properties. This hierarchy ranges from a bottom level with weak "collegiality" (standard tests) to a top level with strong "collegiality," where multiple factors must jointly show an effect.

A key innovation is Property-Oriented Rejection. Similar to a radar system that triggers a specific alarm only when an object moves in a particular direction, this framework allows researchers to define a "direction of interest" (such as a specific risk effect). This directional testing can yield significantly smaller p-values and more precise results than traditional two-tailed approaches.

Handling high-dimensional data requires transforming correlated raw variables into independent intrinsic factors. The framework employs techniques like Factor Analysis (FA) or Non-Gaussian Factor Analysis (NFA) to achieve this "dependence decoupling."

Because testing every possible combination in the lattice is computationally prohibitive, the Best First Path (BFP) algorithm is introduced. BFP efficiently navigates the taxonomy to identify the most distinctive properties without exhaustive traversal. Furthermore, the complex lattice can be "folded" into two layers—one using "AND" logic for joint effects and another using "OR" logic for individual effects—improving interpretability for researchers.

Applications in Genomics and Expression

The framework is particularly suited for biological "data cubes," such as mRNA and lncRNA expression profiles. It enables the identification of biomarkers for cancer stages and survival rates by focusing on joint effects.

Figure 1: Phenotype-targeted testing in a two-variate space. Panel (a) shows the distribution of two groups, while panel (b) demonstrates how a directional vector $\beta$ and decision boundary can separate these groups more effectively than univariate methods.

As shown in the figure above, phenotype-targeted (PT) tests can separate groups that might appear overlapping in a traditional 1D analysis. By considering variables jointly in a multi-dimensional space, the method captures the underlying signal more robustly.

Conclusion

This new multivariate formulation moves beyond "all-or-nothing" testing. By leveraging hierarchical structures and directional analysis, it provides a rigorous way to bound significance and improve sensitivity in genomic studies. While the initial decoupling of factors requires computational effort, the resulting gains in specificity and the ability to handle noisy "do-not-care" variables make it a powerful tool for modern bioinformatics.

Summary updated

Abstract

A new formulation is proposed for multivariate test, consisting of not only a hierarchy of numerous tests organised in a lattice taxonomy of properties that come from different combinations of multi-variates and represent different factors associated with the rejection of null hypothesis, but also by a theory of property-oriented rejection. Located on the bottom level of this taxonomy is a conventional formulation of multivariate test, featured by a property with the weakest collegiality and a rejection with the largest p value. From one level up to the next, the dimension of rejection increases, the collegiality of properties strengthen, and the p values reduce, until the top level that is featured by a property with the strongest collegiality and a rejection with the smallest p value. Instead of traversing all the combinations in the taxonomy, an easy implementation is developed to identify distinctive properties by the best first path (BFP) in a lattice taxonomy of an appropriate number of intrinsic factors that are obtained after decoupling second-order dependence cross multivariate statistics and discarding those non-distinctive components. Even away off this BFP, if needed, a particular combination of intrinsic factors may be conveniently tested in such a taxonomy too. Moreover, further improvement is made by considering some dependence of higher than second order, with the top level p value refined into one upper bound that is obtained by directional test. Furthermore, detailed implementations are also provided for applications to genome-scale sequencing and expression, with particular emphasis on multivariate phenotype-targeted test for expression profile analyses.
BibTeX
@inproceedings{xu2016new,
  author = {Lei Xu},
  title = {{A new multivariate test formulation: theory, implementation, and applications to genome-scale sequencing and expression}},
  booktitle = {Applied Informatics},
  year = {2016},
  doi = {10.1186/s40535-015-0016-4},
}
Figure